Annotation-Efficient Surgical Instance Segmentation via Semi-Supervised YOLO Frameworks on Laparoscopic ImageryShow others and affiliations
2026 (English)In: Proceedings - 2026 13th International Conference on Future Internet of Things and Cloud, FiCloud 2026, Institute of Electrical and Electronics Engineers (IEEE) , 2026, p. 279-285Conference paper, Published paper (Refereed)
Abstract [en]
Accurate instance segmentation of surgical anatomy is essential for computer-assisted surgical systems, but obtaining pixel-wise annotations for laparoscopic images is costly and timeconsuming. This paper investigates annotation-efficient surgical instance segmentation by integrating semi-supervised learning with YOLO-based segmentation models. Using the Dresden Surgical Anatomy Dataset, supervised baselines are first established with YOLOv8-Seg, YOLOv11-Seg, and YOLOv12-Seg. Three semi-supervised strategies, namely pseudo-labeling, MixMatch, and Mean Teacher, are then evaluated under a limited-label setting using YOLOv12-Seg as the base architecture. Experimental results show that all semi-supervised methods improve over the supervised baseline. Pseudo-labeling improves mAP@0.5:0.95 by 11.1%, MixMatch improves it by 27.0%, and Mean Teacher achieves the best result with a 42.9% relative improvement. Qualitative analysis further shows that the Mean Teacher model produces more coherent instance masks under challenging laparoscopic conditions, including occlusion, specular reflection, and illumination variation. These findings indicate that semisupervised learning can reduce dependence on extensive pixelwise annotations while improving surgical anatomy segmentation performance in annotation-constrained settings.
Place, publisher, year, edition, pages
Institute of Electrical and Electronics Engineers (IEEE) , 2026. p. 279-285
Series
International Conference on Future Internet of Things and Cloud (FiCloud), ISSN 2996-1017
Keywords [en]
Surgical vision, instance segmentation, semi supervised learning, YOLO, computer-assisted surgery, laparo scopic imaging
National Category
Computer graphics and computer vision Medical Imaging
Research subject
Cyber Security
Identifiers
URN: urn:nbn:se:ltu:diva-119811DOI: 10.1109/FiCloud70576.2026.00045Scopus ID: 2-s2.0-105049608015OAI: oai:DiVA.org:ltu-119811DiVA, id: diva2:2103438
Conference
13th International Conference on Future Internet of Things and Cloud (FiCloud 2026), Granada, Spain, July 20-22, 2026
Note
ISBN for host publication: 979-8-3195-4739-2
2026-09-212026-09-212026-09-21Bibliographically approved